28-01-2013, 03:40 PM
Application of artificial neural networks for the estimation of tumour characteristics in biological tissues
Application of artificial.pdf (Size: 657.36 KB / Downloads: 25)
Abstract
Background Artificial tactile sensing is a method in which the existence
of tumours in biological tissues can be detected and computerized inverse
analyses used to produce ‘forward results’.
Methods Three feed-forward neural networks (FFNN) have been developed
for the estimation of tumour characteristics. Each network provides
one of the three parameters of the tumour, i.e. diameter, depth and
tumour : tissue stiffness ratio. A resilient back-propagation (RP) algorithm
with a leave-one-out (LOO) cross-validation approach is used for training
purposes.
Results The proposed inverse approach based on neural networks is a
reliable and efficient tool for diagnostic tests in order to accurately estimate
the basic parameters of the tumour in the tissue.
Conclusion There is a non-linear correlation between the tumour
characteristics and their effects on the extracted features. In general, reliable
estimation of tumour stiffness is obtained when the depth of tumour is small.
Copyright 2007 John Wiley & Sons, Ltd.
Introduction
In spite of the huge amount of research activity devoted to the field of breast
cancer study, breast cancer is still the leading cause of death among women
all over the world. As an example, just around 180 000 breast cancers are
diagnosed annually in the USA, and many researchers around the world are
doing theoretical and experimental research to help physicians in saving these
patients (1–3).
The diagnostic accuracy of tumour parameters is a special concern
in breast specialist units when further treatment is being planned. The
decision for or against primary surgery vs. primary neo-adjuvant chemotherapy
depends mostly on the initial diagnostic staging, including tumour
size, depth and stiffness (4). Due to the diagnostic uncertainty that
results from screening, physicians must use a variety of diagnostic tests,
such as supplemental mammography, physical examination, fine needle
aspiration cytology, ultrasound, gadolinium-enhanced magnetic resonance
imaging (MRI) and/or positron emission tomography (PET) (5).
Background and definition of the
problem
In many diagnostic tests, such as clinical breast
examinations, doctors routinely examine the patient’s
body with the fingers and palm to obtain information
on conditions inside the body, including the presence
of a tumour and its precise characteristics. In actual
practice, when a woman visits her physician, part of the
physical examination often includes a clinician attempting
to palpate the patient for any lumps or changes in the
breast tissue that could indicate the presence of a tumour.
This method, however, only gives the physician a vague
sense of what is actually underneath the skin. Due to the
lack of any precise and consistent measuring approach, if
a lump is found through palpation, typically all that can
be documented is its general location on the breast and a
rough estimate of its size. Moreover, if the patient visits
other clinicians, she will undoubtedly receive different
descriptions of her status. This is because when there is
no consistent method, there is no consistent description. In
the present studywe developed the concept of a consistent
approach which neither needs the dexterity of a reputable
physician nor has any penetration/invasion into the body,
such as the effects of X-ray in mammography.
Materials and methods
An understanding of the biological mechanism of the
human nervous system and its structure has significant
effects on the design of ANNs. In the conventional
structure of an ANN, the weights correspond to the
synapses in a biological neuron, while the activation
function is associated with the intracellular current
conduction mechanism in the soma. An artificial neuron
is an oversimplified but useful approximation of the
biological neuron. This simple model ignores many of
the characteristics of its biological counterpart, e.g.
Simulations and results
In the supervised learning paradigm, a set of experimental
pairs (forward analyses) of an input–output mapping is
needed to train the neural network. In other words, we
wish to deduce the mapping implied by the data. Thus,
the cost function is related to the mismatch between our
mapping and the data. The training data samples have
to be fairly large to contain all the required information
and must include a wide variety of data from different
experimental conditions and process parameters (32). In
this work, we used 30 data samples obtained from the
forward modelling to train ANN. A data sample was a
vector with five elements (numbers), which were the
extracted features of a stress graph.
Discussion and conclusions
Considering the diagnostic tests carried out by physicians
to detect tumours in the patient’s body and the forward
analyses done in our previous work (described above),
the ANN approach was employed to determine the
tumour characteristics, i.e. diameter, depth and stiffness.
Two neural networks were implemented to estimate
the tumour diameter and depth, respectively; however,
estimation of the stiffness ratio (Er) was more difficult
than the other tumour characteristics, since the stress
graphs did not provide enough information. This difficulty
was remedied by estimating Er in two stages. At the first
stage h and d were estimated, and next, based on these
estimations and the corresponding maximum values of the
stress graphs, Er was estimated. Training the networks
was achieved using a back-propagation algorithm with
a LOO cross-validation approach.